You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@drill.apache.org by "Paul Rogers (JIRA)" <ji...@apache.org> on 2017/03/20 21:50:41 UTC

[jira] [Created] (DRILL-5370) Drillbit dies for 5 MB SELECT statement

Paul Rogers created DRILL-5370:
----------------------------------

             Summary: Drillbit dies for 5 MB SELECT statement
                 Key: DRILL-5370
                 URL: https://issues.apache.org/jira/browse/DRILL-5370
             Project: Apache Drill
          Issue Type: Bug
    Affects Versions: 1.10.0
            Reporter: Paul Rogers


Some community users use Drill with BI tools that generate queries. One such tool generates queries that map Drill data into a "cube" format for a cube-based visualization engine. Such tools tend to create very large, very complex queries.

In replicating an issue found by this user, I created a simple program that creates deeply-nested queries of the form:

SELECT a99 AS a98 FROM (SELECT a97 AS a98 FROM(… SELECT a1 FROM myTable)…))

The test used 200 columns each with names of 500 characters long. (Drill has a hard limit of 1024 characters for a symbol name.)

The setup was an embedded Drillbit using the new "cluster fixture" test framework. The test ran multiple iterations, each wrapping the prior SELECT in a new one as shown above. The result is a series of queries that grew in size by about 100K each iteration.

Drill handled SELECT statements up to 5 MB in size, after which the Drillbit ran out of heap memory, suffered a fatal exception and exited.

One question is why a 5 MB query exhausted multiple GB of heap during query parsing and planning.

But, more importantly, Drill should have some way to protect itself from such failures. In a production cluster, heap exhaustion will bring down all in-flight queries and require a manual restart of the Drillbit.

So, Drill should enforce some limit on the amount of heap memory used by a query during the parsing and planning process.

The community user found a failure at around 1 MB, but they very likely had a query with much more complex structure than the simple nested-SELECT used in my test.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)